does fertility decrease household consumption? an analysis ... · in addition to joint causation,...

Demographic Research a free, expedited, online journalof peer-reviewed research and commentaryin the population sciences published by theMax Planck Institute for Demographic ResearchKonrad-Zuse Str. 1, D-18057 Rostock · GERMANYwww.demographic-research.org

DEMOGRAPHIC RESEARCH

VOLUME 20, ARTICLE 26, PAGES 623-656PUBLISHED 05 JUNE 2009http://www.demographic-research.org/Volumes/Vol20/26/DOI: 10.4054/DemRes.2009.20.26

Research Article

Does fertility decrease household consumption?An analysis of poverty dynamics and fertility inIndonesia

Jungho Kim

Henriette Engelhardt

Alexia Prskawetz

Arnstein Aassve

c© 2009 Jungho Kim et al.

This open-access work is published under the terms of the CreativeCommons Attribution NonCommercial License 2.0 Germany, which permitsuse, reproduction & distribution in any medium for non-commercialpurposes, provided the original author(s) and source are given credit.See http://creativecommons.org/licenses/by-nc/2.0/de/

Table of Contents1 Introduction 624

2 Measuring household consumption 626

3 Statistical methods 631

4 Institutional background and data 6344.1 Background and data description 6344.2 Control variables 635

5 Household consumption expenditure and fertility 638

6 Private expenditure on child and intra-household bargaining 643

7 Conclusion 650

8 Acknowledgements 651

References 652

Appendix: Estimating Equivalence Scale using Engel method 655

Demographic Research: Volume 20, Article 26

Research Article

Does fertility decrease household consumption?An analysis of poverty dynamics and fertility in Indonesia

Jungho Kim 1

Henriette Engelhardt 2

Alexia Prskawetz 3

Arnstein Aassve 4

Abstract

This paper presents an empirical analysis of the relationship between fertility and povertyfor Indonesia, a country which has experienced unprecedented economic growth andsharp fertility decline over recent decades. We illustrate the sensitivity of the effect offertility on household consumption with respect to the equivalence scale in a unitaryhousehold framework. Using the propensity score matching method, the analysis sug-gests that a newborn child decreases household consumption per person by 20 percentwithin four years. When the estimates of equivalence scales implied by the Indonesiansample are applied, the effect of a child on household consumption is still negative, butthe magnitudes are in the range of 20 to 65 percent of that found with the per-capita ex-penditure as a measure of consumption. Therefore, it is suggested that analysis basedon the conventional measure of poverty is likely to exaggerate the effect of fertility onpoverty at least because of neglect of the proper equivalence scale. Given that householdpreference for consumption of private goods such as children’s education is negativelyassociated with fertility, the test for household bargaining supports the model of the uni-tary household as a valid assumption for examining the relationship between fertility andhousehold consumption.

1 Korea Development Institute, Cheongnyangni-dong Dongdaemun-gu, Seoul 130-740, Korea,Phone: +82 2-958-4742, Fax: +82 2-958-4090, E-mail: [email protected] Otto-Friedrich-Universität Bamberg, Professur für Bevölkerungswissenschaft, Lichtenhaidestraße 11,D-96045 Bamberg, Germany, Phone: +49 0951-863-2645, Fax: +49 0951-863-5644.E-mail: [email protected] Institute for Mathematical Methods in Economics (Research Unit Economics), Vienna University of Tech-nology, Argentinierstr. 8/4/105-3, 1040 Vienna, Austria and Vienna Institute of Demography, Wohlleben-gasse 12-14, 6th floor, 1040 Vienna, Austria. Phone: +43 1-58801-17510. Fax: +43 1-58801-17599.E-mail: [email protected] Universita Bocconi, Viale Isonzo 25, Department of Decision Sciences, 20135 Milano, Italy, Phone: +3902-5836-5657, Fax: +39 02-5836-5630, E-mail: [email protected].

http://www.demographic-research.org 623

Kim et al.: Does fertility decrease household consumption?

1. Introduction

The causal relationship between population growth and standard of living has long beenof interest to policymakers. While the estimates of the effect of fertility on poverty rangefrom being significantly positive to insignificant, a growing body of literature suggeststhat the true relationship is likely to be more complicated than unidirectional (e.g. Bird-sall et al. 2001). The relationship between poverty and fertility is not unidirectionalbut dependent on the stage of economic development (McNicoll 1997, Schoumaker andTabutin 1999). While in most contemporary developing countries this relationship ispositive, a negative relationship has been reported within the poorest countries. The lat-ter results are associated with lower reproduction capability and higher rates of infertilityamong extremely poor households (Lipton 1998, Livi-Bacci and De Santis 1998). Clearlymany factors that influence fertility also determine well-being. These include education,health services and family planning policies. In addition to joint causation, reverse causa-tion may also take place. Among poor households, the demand for children is high sincethose households rely on their children’s labor supply and often the child’s support is crit-ical when parents become old. Higher fertility in turn is associated with lower investmentin education (i.e. demand for quantity rather than quality of children) and consequentlylower earnings potential for children, which in turn fosters intergenerational transmissionof poverty (Moav 2005).

The theoretical perspectives suggest various mechanisms linking fertility and poverty.Empirical studies that tried to identify the causal relationship between fertility and povertyhave so far relied on aggregate level data and cross-sectional micro level data (cf. the re-view by Merrick 2001). With these data, it is difficult to provide robust causal informationabout fertility and well-being because fertility and household income are jointly deter-mined. Recent longitudinal household surveys in developing countries that incorporatethe timing of fertility together with information on consumption expenditure, income andother measures of well-being allow researchers to identify the dynamic relation betweenpoverty and fertility. So far, these data sets have not been used to study the link betweenpoverty and fertility, which is the goal of this paper. In particular, we examine the Indone-sian experience for which we have excellent longitudinal information on both fertility andhousehold expenditure, together with a range of other background information.

The most commonly used measure of poverty is a dichotomous variable that indi-cates whether a household’s consumption expenditure is below a certain poverty thresh-old level. Although this poverty measure is useful in comparing the standard of livingacross countries or over time, its analysis cannot measure the magnitude of the effect offertility on household consumption. Moreover, the thresholds are often set in an arbitraryway. Therefore, we examine the relationship between fertility and household consump-tion directly. It should be noted that the level of consumption is not the only measure

624 http://www.demographic-research.org


of household welfare. In addition, parents may enjoy having children and children maymake the consumption of other goods more enjoyable. However, in light of the Mil-lennium Development Goals, set out by the United Nations, the materialistic measure ofwelfare is clearly of interest.5 As a result, we examine here whether and how much havingan additional child decreases household consumption.

One difficulty in establishing a relationship between fertility and household consump-tion is related to the measure of consumption (Ravallion 1996). As argued in Lanjouw andRavallion (1995) “Despite extensive work on welfare measurement in economics, thereis still no preferred method for making inter-personal comparisons across households ofdifferent size and/or composition.” For instance, a common measure of well-being at thehousehold level is expenditure per adult equivalent. However, assigning equal weight toall household members and overlooking economy of scale with household size ignoresthe compositional and size effect of households on the measure of well-being. As shownin Lanjouw and Ravallion (1995) for Pakistan, the measure of poverty is sensitive to theeconomy of scale parameter that is assumed. In this paper we investigate whether thecausal relation between fertility and household consumption is sensitive to two dimen-sions of equivalence scale frequently discussed in the literature: the weight of a child’sconsumption relative to an adult’s and the economy of scale (Banks and Johnson 1994;Koulovatianos et al. 2005)

A further complication in verifying the relationship between fertility and poverty is theissue of measuring a variation of fertility that is exogenous with respect to the measuresof well-being. Two approaches have been adopted in the existing literature. The first oneis to use the context of a natural experiment such as samples of twins or the sex composi-tion of the first two children as an instrumental variable for fertility (e.g. Rosenzweig andWolpin 1980; Angrist and Evans 1998). Although these instruments are reasonably valid,Rosenzweig and Wolpin (2000) note that this approach requires restrictions on the utilityfunction (e.g. the separability of leisure and consumption, etc.). The second approach is touse the residual from a fertility regression as a measure of unobserved fecundity (Rosen-zweig and Schultz 1985). Measuring female fecundity is certainly a useful approach, butthe result is less intuitive. As an alternative approach, we apply here a matching method inorder to estimate the effect of an additional child on household consumption in the shortrun. This approach assumes that the event of childbearing is independent of householdconsumption given observable characteristics. However, our focus is on illustrating theextent to which the effect of fertility on poverty depends on the measure of equivalencescale under a reasonable estimation strategy, and we minimize any possible bias due to

5 The first goal of the Millennium Development Goal is to eradicate extreme poverty and hunger by 2015. Ithas two specific objects. The first is to reduce by half the proportion of people living on less than a dollar aday, and the other is to reduce by half the proportion of people who suffer from hunger. The full description isavailable at http://www.un.org/millenniumgoals/.



unobservable characteristics by conditioning on an extensive set of observable character-istics.6

The sensitivity of the effect of fertility on household consumption suggests that theexpenditure share of private good may affect fertility decision. Then, the share of pri-vate good in household expenditure may be determined by the intrahousehold bargainingprocess. Therefore, these possibilities are further investigated in order to find an empiricalevidence that connects unitary and collective models of a household.

Our main finding is that households with a new born child between 1993 and 1997experience about 20 percent reduction in consumption when per-capita consumption isused as a measure of household consumption. However, the effect of fertility on con-sumption is highly sensitive to the choice of the equivalence scale. When we apply theestimated equivalence scale parameters with further assumptions on household welfare,the magnitude of the effect of fertility on consumption is in the range 20 to 65 percent ofthat found with the per-capita consumption. Therefore, it is suggested that the analysisbased on the conventional measure of poverty is likely to exaggerate the effect of fertilityon poverty at least because of the neglect of the proper equivalence scale. On the otherhand, the household preference for consumption of private goods such as children’s edu-cation is found to be negatively correlated with fertility. Further, the results of the tests forthe household bargaining process supports the framework of a unitary household model,which is a basic assumption for the analysis of equivalence scale.

The rest of the paper is organized as follows. Section 2 introduces the measure ofhousehold consumption and summarizes estimates of the equivalence scale implied bythe data. The method of analysis is discussed in section 3. Section 4 describes the in-stitutional background, data and variables used in our study. The results on whether andin which direction fertility causes household consumption are presented in section 5. Weapply household consumption expenditure per person and investigate the sensitivity ofthe results depending on the way in which household size and household compositionare taken care of. Section 6 examines the effect of private goods share of expenditureon fertility and the effect of bargaining power on the share of private goods in householdexpenditure. In section 7 we conclude.

2. Measuring household consumption

The most commonly used measure of welfare is household expenditure per person. Acritical issue when considering the effect of a new born child on household consumptionis the definition of the equivalence scale (Lanjouw and Ravallion 1995). One approachis to directly estimate the cost of a child in terms of an adult’s consumption (e.g. Deaton

6 See Mattei (2004) for a similar approach.



and Muellbauer 1986). A limitation of this approach is that the calculation is based onhouseholds of a specific demographic type. Therefore, the result is not directly applicableto households of various demographic types. Unlike developed countries where nuclearhouseholds are common, there are 67 different types of households in our sample withthe numbers of children and adults ranging from zero to ten and from one to twelve, re-spectively. Although, in principle, it is correct to calculate the cost of a child for eachhousehold type separately, it seems more appropriate here to estimate the effect of a new-born child in an average household.

One way of estimating the equivalence scale is to use a complete demand system andto estimate the equivalence scale for a particular type of household compared to a base-line household (e.g. De Santis and Maltagliani 2003). Our approach is to estimate twoparameters of equivalence scale in a functional form, which can be applied to all house-holds with different demographic composition. It has the disadvantage of assuming aparticular functional form for the equivalence scale, but, at the same time, has an impor-tant advantage since it allows us to evaluate the effect of a birth on individual householdconsumption, on average, under the equivalence scale implied by the data. The functionalform is widely used, and in this sense we are in line with the majority of the previousliterature.

There are two dimensions of the equivalence scale to be considered: 1) the weightassigned to children relative to adults, and 2) the economy of scale in household con-sumption. We adopt a simple form of a welfare measure incorporating these dimensions,as suggested by Banks and Johnson (1994).7

W =H

(A + αK)θ, (1)

where H denotes total household expenditure, and A and K denote the number of adultsand the number of children respectively. A measure of household consumption per equiv-alence scale, W , is the measure of household consumption when the cost of each memberis taken into account. Therefore, if two households of different demographic character-istics have the same level of W , then it can be argued that they have the same level ofwelfare. In equation (1), the weight for a child relative to an adult is α, and the economyof scale is reflected through the parameter θ. Both α and θ take values between zeroand one. Using the expenditure per person as a measure of household consumption im-plies that a child consumes as much as an adult and that there is no economy of scale inconsumption of goods (α = 1, θ = 1).

7 We are aware that the functional form used in the paper is not the only one used, but this version is a frequentlyused specification in the literature (e.g. Banks and Johnson 1994; Koulovatianos et al. 2005).



Since estimating the equivalence scale requires a set of assumptions about the householdpreference structure, we first study the effect of a child on consumption expenditure fordifferent values of the equivalence scale (α = 1/2, 1; θ = 1/2, 1), and then apply ourestimates of equivalence scale to illustrate the range of the effect of a child on individualconsumption.

One way to estimate α and θ in equation (1) is to use experimental data as in Koulo-vatianos et al. (2005). Given the lack of the data on direct compensation for an additionalchild or an adult, we take an alternative approach using the Engel curve estimation as inDeaton and Muellbauer (1986).8 That is, under the assumption that the inverse of foodshare of expenditure correctly indicates the welfare of household members, we calcu-late the amount of consumption that is needed in order to compensate for the additionalmembers of the household compared to the baseline household. A household with twoadults is taken as the baseline, and the estimated equivalence scales for households withdifferent demographic characteristics are presented in Table 1. The estimation procedureis described in Appendix A. The second row in Table 1 states that a household with twoadults and a child needs 126 percent of the consumption of the reference household inorder to have the same level of welfare, suggesting that a child costs 52 percent of anadult’s consumption.9

It is notable that the estimation procedure taken in the paper considers food as havinga private good aspect. This assumption is not against the finding by Deaton and Paxson(1998). They show that conditional on the per capita expenditure the food expenditureper person as a private good should increase as household size increases but that datasuggest the opposite. Since food having a public good aspect did not prove to be a goodexplanation for their puzzle, our assumption seems to be reasonable with regard to theirfindings.

As Blundell and Lewbel (1991) pointed out, the conditional demand equation can beused to identify the cost of living indices for different household compositions but not thetrue equivalence scale. Restoring the equivalence scale requires an identifying assumptionsuch as independence of base utility, or the assumption that unconditional preferenceorderings depend on demographics only through Barten scales. With this limitation, weuse the relative magnitudes of the costs of additional household members in order toestimate the parameters of equivalence scale.

8 The Engel curve refers to the demand for a good (food in this case) as a function of income when the pricesare fixed.9 Using an Indonesian survey (Susenas) in 1978, Deaton and Muellbauer (1986) estimated that a child costs 90percent of an adult’s consumption using Engel’s method.



Table 1: Equivalence scale for different household types

Scale(S) Household type No. adults (A) No. children (K)1.00 AA (baseline) 2 01.26 AAK 2 11.41 AAKK 2 21.48 AAKKK 2 31.16 AAA 3 01.30 AAAK 3 11.37 AAAKK 3 21.38 AAAKKK 3 3

Notes: 1) Data source: The 1993 Indonesian Family Life Survey.2) A child is defined as a household member under age 15.3) The estimation is conducted under the assumption that the inverse of food share of expenditure cor-rectly indicates the welfare of household members.4) The estimated scale indicates the amount of consumption (compared to that of the baseline household)needed to make the household as well-off as the baseline household.

Given that there is a wide variety of household types in the sample, we run the fol-lowing non-linear regression to find a set of parameters of equivalence scale, using theequivalence scales for eight different household types as shown in Table 1. We do thisinstead of estimating the equivalence scale for every household type separately.

S = δ(A + αK)θ + ε, (2)

where S represents the estimated total equivalence scale, and δ is an additional scaleparameter. Because we have only eight observations for estimating the parameters ofthe equivalence scale function, the estimates are often out of the range implied by thetheory. By introducing additional parameter of δ we search for the range of estimates ofα and θ satisfying the theoretical prediction. Although δ is not identified, it should benoted that the parameter δ does not affect the comparison of two households of differentdemographic types. The purpose of our study is not to estimate the exact equivalence scalefor a specific practical application (such as a tax policy) but to find a range of reasonableestimates of α and θ that is implied by the data. The estimation procedure is that the valuesof α and θ are estimated for a fixed value of δ. Then, we find a range of δ that producesthe estimates of α and θ between zero and one as indicated in Table 2. In the followingsections, we estimate the effect of a newly born child on household consumption perequivalence scale for the range of the estimates in Table 2.



Table 2: Estimated parameters of equivalence scale

δ α̂ t-ratio (α̂) θ̂ t-ratio (θ̂) Residual R2

0.1 0.475 (1.28) 1.990 (6.00) 1.609 0.8820.2 0.525 (1.43) 1.454 (6.32) 0.901 0.9340.3 0.581 (1.55) 1.132 (6.53) 0.547 0.9600.4 0.654 (1.68) 0.898 (6.70) 0.336 0.9750.5 0.757 (1.81) 0.712 (6.83) 0.203 0.9850.6 0.916 (1.95) 0.556 (6.93) 0.118 0.9910.7 1.192 (2.07) 0.420 (6.93) 0.064 0.9950.8 1.780 (2.10) 0.298 (6.65) 0.032 0.9980.9 3.527 (1.77) 0.188 (5.61) 0.016 0.9991.0 10.343 (0.94) 0.102 (3.37) 0.015 0.9991.1 53.709 (0.26) 0.047 (1.26) 0.033 0.9981.2 4, 702, 439.000 . 0.008 (4.29) 0.072 0.5661.3 −0.617 (7.28) −0.078 (1.37) 0.094 0.9931.4 −0.477 (5.06) −0.153 (2.88) 0.099 0.9931.5 −0.363 (3.39) −0.227 (3.82) 0.121 0.9911.6 −0.284 (2.38) −0.294 (4.38) 0.149 0.9891.7 −0.226 (1.76) −0.356 (4.75) 0.181 0.9871.8 −0.183 (1.34) −0.415 (5.01) 0.215 0.9841.9 −0.149 (1.05) −0.470 (5.20) 0.250 0.9822.0 −0.122 (0.83) −0.523 (5.35) 0.288 0.979

Notes: 1) Data source: Table 1.2) The estimation is conducted under the assumption that the inverse of food share of expenditure cor-rectly indicates the welfare of household members.



3. Statistical methods

In order to estimate the causal effect of fertility on material wellbeing, we apply a match-ing approach based on the treatment effect literature following the counterfactual modelof causal inference. This approach is based on the intuitively attractive idea of contrast-ing the outcomes of a treatment group, Y1, with the outcomes of a ‘comparable’ controlgroup, Y0, conditional on a set of characteristics of individuals, X . The outcome in thisstudy is the household consumption expenditure per equivalence scale in 1997, and thetreatment is defined as the birth of a child between 1993 and 1997. The control variable,X , includes the observable characteristics of households in 1993. Differences in the out-comes between the two groups are consequently attributed to the treatment, D. Matchingmethods for estimating causal effects have several advantages. First, they make no as-sumptions about the functional form of the dependence between the outcome of interestand the control variables. Second, matching ensures that the control variables of interestin the treatment group are similar to those in the control group and, thus, only similar unitsare compared. Third, since fewer parameters are estimated than in a traditional regressionmodel, matching may be more efficient. This can be especially important if samples aresmall.

The matching method is based on the identifying assumption that, conditional on X ,the outcome of the control group Y0 is independent of the treatment D. Using the notationof Dawid (1979), the assumption of strongly ignorable treatment assignment

Y0 ⊥⊥ D | X, (3)

is sufficient to identify the mean effect of treatment on the treated, or ATT (Rosenbaumand Rubin 1983):

ATT = E(Y1 − Y0 | D = 1) = E(Y1 | D = 1)− E(Y0 | D = 0). (4)

Assumption (3) produces a comparison group that resembles the control group of an ex-periment in one key aspect: conditional on X , the distribution of Y0 given D = 1 is thesame as the distribution of Y0 given D = 0:

E(Y0 | X, D = 1) = E(Y0 | X, D = 0) = E(Y0 | X). (5)



In general, the households that experience the event of childbearing are likely to be differ-ent from those who do not in many respects. Examples include educational achievement,working status, preference for having children, and so forth. Thus a simple tabulationof the treatment and controls will most likely yield biased estimates of the treatment ef-fect. In other words, these differences are going to affect the total household consumptionexpenditure, which induces a bias of the estimate due to selection. By matching thehouseholds based on a set of control variables, ATT in equation (4) corrects for the selec-tion bias due to the correlation between childbearing and the observable characteristicsof households (Rosenbaum and Rubin 1983). Strictly speaking, the matching methoddoes not resolve any selection bias due to the unobservable characteristics (Heckman andRobb 1985). However, in so far as the observed variables used for matching also cap-ture outcomes due to unobserved heterogeneity, this bias will be reduced. Hence, the setof variables included in the matching should be as extensive as possible, including bothexogenous and endogenous variables. To further refine our estimates, we consider thedifference between outcome variables in 1993 and 1997.

ATT = E((Y1,1997 − Y1,1993)− (Y0,1997 − Y0,1993) | D = 1)= E(Y1,1997 − Y1,1993 | D = 1)− E(Y0,1997 − Y0,1993 | D = 0)= E(∆Y1 | D = 1)− E(∆Y0 | D = 0). (6)

The ATT in equation (6) is the so-called difference-in-difference estimator, and correctsfor the selection bias due to the fixed component of unobservable characteristics. Thatis, if the effect of the unobservable characteristics of a household such as the preferencefor childbearing on individual household consumption is present in both 1993 and 1997and if its magnitudes are the same, it is removed in equation (6) as the change in theoutcome variable is considered. It should be noted that the identifying assumption forATT in equation (6) has a weaker form implying that the change in the outcome of thecontrol group is independent of the treatment.

Y0,1997 − Y0,1993 ⊥⊥ D | X. (7)

However, our analysis is still prone to selection bias due to the time-varying componentof the unobservable characteristics.

For estimating the mean effect of treatment on the treated, many matching estimatorshave been proposed that exploit assumption (7) for alternative matching methods. Forall matching methods, the average treatment effect for the treated in equation (6) can bewritten as (Heckman et al. 1998):



E(∆Y1 −∆Y0 | D = 1) =∑

i∈T

wi

∆Y1i −

∑

j∈C

Wi,j∆Y0j

, (8)

that is, the average (weighted by wi) of the differences between the events, Y1i, of thetreatment group T and the events, Y0j , of the control group C weighted by Wi,j . Thedifferent matching algorithms differ in the construction of the comparison weights, Wi,j .

Traditional matching methods pair the non-treated with the treated persons that are‘close’ in terms of X using different metrics, e.g. caliper matching of different widths,Mahlanobis distance matching, or kernel-based matching. In practice (i.e. with samplesof typical size) it is often difficult to match on high dimensional X . Instead it is easierto do the matching based on individuals’ probability of treatment, or in other words thepropensity score (Rosenbaum and Rubin 1983). We define P (X) as the propensity scorewith P (X) = P (T = 1 | X). If the balancing property

D ⊥⊥ X | P (X) (9)

is satisfied, i.e. X and D are independent conditional on P (X), observations with thesame propensity score must have the same distribution of observable and unobservablecharacteristics independent of treatment status. In other words, for a given propensityscore, exposure to treatment is random. A theorem of Rosenbaum and Rubin (1983)demonstrates that if assumption (7) is satisfied, then

Y0,1997 − Y0,1993 ⊥⊥ D | P (X), (10)

provided 0 < P (X) < 1, so that there is a positive probability that the events D = 1and D = 0 occur. This insight shows that matching can be performed on P (X) alone,provided that the balancing property (9) holds.

Propensity scores are implemented in the matching techniques by defining the close-ness of propensity scores and the control variable X in different ways. Since estimatesare sometimes sensitive to the choice of matching technique, we implement several ap-proaches. In particular we apply nearest neighbor matching, radius matching, kernelmatching, and stratification matching based on the propensity score (Becker and Ichino2002). With nearest neighbor matching, each member of the treatment group is matchedto a non-treated unit using the closest propensity score. With radius matching, the treatedunits are only matched with non-treated units within a pre-specified range around theirpropensity scores. With kernel matching, the propensity score of each treated unit ismatched with the kernel-weighted average outcome of all non-treated units.



Finally, with stratification matching, the range of variation of the propensity score is di-vided into intervals such that within each interval, treated and control units have on aver-age the same propensity score.

4. Institutional background and data

4.1 Background and data description

In many developing countries we observe that fertility declines often come along with re-ductions in poverty. Indonesia is perhaps the most striking example of this pattern. Overthe last four decades, Indonesia has experienced unprecedented economic growth togetherwith a dramatic fertility decline. Table 3 shows that the real GDP per person increased bymore than three times over the period 1970 to 1995, and total fertility rate fell by around50% over the same period. This dynamic nature of the socio-economic change combinedwith a large population and its vast geographical diversity has attracted considerable inter-est among demographers and policymakers alike. We focus on the period between 1993and 1997, which is at the end of the fertility transition and before the Asian FinancialCrisis.10

Table 3: Total fertility rate and GDP per capita in Indonesia

Period TFR GDP per capita1965-1970 5.57 297.61970-1975 5.20 384.31975-1980 4.73 503.01980-1985 4.11 601.71985-1990 3.50 776.71990-1995 3.00 1, 048.7

Notes: 1) Source of TFR: World Population Prospects: The 2000 Revision, Vol. I, United Nations PopulationDivision (requoted from World Resources Institute)2) Source of GDP per capita: World Development Indicators 2004, The World Bank.3) GDP per capita is in constant 1995 US dollars, and indicates the value in the last year of each period.

10 The field work of the second wave of the Indonesian Family Life Survey (IFLS2) was conducted betweenAugust and December of 1997. Although there had been a modest increase in the exchange rate for that period,the real financial crisis took off in January 1998. Frankenberg et al. (2003) also point out that the growth rate ofprices and wages in Indonesia started to jump in January 1998. Therefore, it seems reasonable to assume thatthe IFLS2 was not severely affected by the financial crisis.



Our study is based on data from the Indonesian Family and Life Survey (IFLS). TheIFLS consists of three waves in total, IFLS1 conducted in 1993/94, IFLS2 and IFLS2+ in1997 and 1998, and IFLS3 in 2000. In 1997, the Indonesian economy was hit by the Asianfinancial crisis. The second wave of the IFLS was conducted shortly before the event. Wedo not use any of the waves after 1997 in order to prevent estimates from being drivenby the financial crisis. The IFLS is of exceptional quality and is ideal for the purposeof constructing a quasi-experiment of the type implemented here, as the survey containsextensive questionnaires on a range of aspects of household economy covering a period ofsteady economic growth. The survey has been conducted by RAND Corporation in col-laboration with UCLA and Lembaga Demografi, University of Indonesia (Frankenbergand Karoly 1995). The sample is representative of about 83% of the Indonesian popula-tion and contains over 30,000 individuals living in 13 of the 27 provinces in the country.IFLS1 has 7,224 households, and subsequent waves targeted all the split-off householdsas well as all the original households previously interviewed. The response rate for theIFLS2 was 94 percent of the original sample (Frankenberg and Thomas 2000). The surveycontains a wealth of information collected at the individual and household levels, includ-ing multiple indicators of economic well-being such as consumption, income, and assets.It also includes information on education, migration, labor market outcomes, marriage,fertility, contraceptive use and health. Information on relationship among co-residentsand non-co-resident family members and inter-generational mobility are included as well.Another outstanding feature of the IFLS is the quality of information provided at the com-munity level. The panel has information concerning the physical and social environment,infrastructure, employment opportunities, food prices, access to health and educationalfacilities, and the quality and prices of services available at those facilities.

4.2 Control variables

Our unit of analysis is the household. 6,742 households were interviewed in both waves.After dropping observations that have missing values for relevant variables, the final sam-ple consists of 4,852 households. Table 4 presents summary statistics of household char-acteristics. The proportion of households with urban residence is 44% in 1993 and re-mains almost the same in 1997. Average household size is 4.57 in 1993 and it decreasesby 0.1 in 1997. The proportion of children within households decreases from 0.31 to0.28 between waves. The average number of newly born children between waves is 0.31.Put differently, 27% of the households experienced at least one live birth. Only 15% ofhouseholds are headed by a female, and the head’s average age is 45.6 years in 1993 and48.4 in 1997. Islam is the dominant religion, as 87% of household heads are Muslim. Theproportion of household heads who worked in the previous year increases from 82% to86% between waves.



The educational attainment of household heads increases from 4.7 years in 1993 to 5.3years in 1997. The proportion of households with new household heads in the secondwave is four percent. General educational attainment of household members increasesbetween 1993 and 1997 as can be seen in Table 2. The average number of adult menwho worked in the previous year increases by 0.05, but that for adult women decreasesby 0.07 in the second wave. The prevalence of child labor is low in the sample, whichis partly due to the fact that the question of whether an individual worked in the pastyear does not capture part-time workers or unpaid family workers. The proportion ofhouseholds with at least one farmer decreases from 42% in 1993 to 37% in 1997. Realmonthly expenditure per person is 39,106 Rupiah in 1993 (in 1986 Rupiah), and it slightlydecreases to 38,808 in 1997 (in 1986 Rupiah).11 The expenditure share of food is 61% in1993, and it increases to 67% in 1997.

Table 4: Summary statistics of the balanced panel (N = 4,852)

1993 1997Variable Mean S. D. Mean S. D.

Demographic variablesUrban residence (index) 0.44 (0.50) 0.43 (0.50)Household size 4.57 (2.07) 4.47 (2.01)Proportion of children in household 0.31 (0.22) 0.28 (0.22)Proportion of adults in household 0.69 (0.22) 0.72 (0.22)No. of children 1.62 (1.41) 1.44 (1.32)No. of adults 2.95 (1.39) 3.03 (1.43)No. of head’s own children 2.13 (1.66) 2.01 (1.56)No. of head’s own sons 1.11 (1.15) 1.05 (1.09)No. of head’s own daughters 1.02 (1.06) 0.96 (1.01)No. of generations 2.09 (0.58) 2.16 (0.60)No. of newly born children between waves 0.31 (0.55)

11 The consumption expenditure includes expenditures on food, non-food and education. The questionnaireof the IFLS asks about food expenditure in the past week, non-food expenditure in the past month (for thenon-durable) or in the past year (for the durable), and educational expenditure in the past year.



Table 4: (Continued)

1993 1997Variable Mean S. D. Mean S. D.

Household head characteristicsHead is female (index) 0.15 (0.36) 0.17 (0.38)Head’s age 45.57 (14.21) 48.38 (13.81)Head is Muslim (index) 0.87 (0.34) 0.87 (0.34)Head worked last year (index) 0.82 (0.38) 0.86 (0.35)Head’s years of schooling 4.74 (4.22) 5.34 (4.36)Head is married (index) 0.85 (0.35) 0.83 (0.38)Head has a spouse in HH (index) 0.81 (0.39) 0.79 (0.41)Head is new in the second wave (index) 0.04 (0.19)

Educational attainmentNo. of adult men with more than primary education 0.55 (0.81) 0.63 (0.86)No. of adult women with more than primary education 0.43 (0.72) 0.53 (0.79)No. of adult men with only primary education 0.30 (0.56) 0.33 (0.56)No. of adult women with only primary education 0.29 (0.51) 0.34 (0.54)No. of adult men with less than primary education 0.57 (0.66) 0.47 (0.63)No. of adult women with less than primary education 0.82 (0.71) 0.72 (0.68)

Working statusNo. of adult men who worked last year 1.01 (0.70) 1.06 (0.67)No. of adult women who worked last year 0.69 (0.70) 0.62 (0.68)No. of male children who worked last year 0.02 (0.16) 0.01 (0.09)No. of female children who worked last year 0.02 (0.14) 0.00 (0.07)Household with at least one farmer (index) 0.42 (0.49) 0.37 (0.48)

Measure of welfareReal monthly expenditure per person/100 (1986 Rupiah) 391.06 (324.99) 388.08 (308.12)Expenditure share of food 0.61 (0.20) 0.67 (0.18)

Notes: 1) Data source: IFLS1 (1993) and IFLS2 (1997).2) 21 observations are dropped due to missing expenditure share of food.3) Only primary education (six years of schooling) was compulsory until 1994. Since then, junior highschool became mandatory. However, the full enrollment of junior high school has not been achieved.4) A child is defined as a household member who is less than 15 years old.5) The sampling weight is not applied.



5. Household consumption expenditure and fertility

This section presents our estimates of the effect of a new born child on household con-sumption using the average treatment effect model. The estimation is conducted in twostages. The first stage involves estimating the probability of receiving the treatment, inthis case, having a child born between two waves is estimated. In the second stage, theaverage treatment effect is estimated by comparing households with and without the treat-ment that have similar propensity scores.

Since the main purpose for probit estimation is to construct a set of variables that canbe used as a basis for matching households, the specification includes an extensive set ofvariables representing a household’s demographic characteristics, educational attainmentand working status including the head’s characteristics. However, there still remains anissue of unobservable characteristics such as cultural norms that might affect female la-bor force participation and fertility. The identifying assumption for the matching methodis that the treatment (i.e. to have a newborn child between 1993 and 1997) is indepen-dent of the characteristics observed in 1993 including female working status and femaleeducation. This assumption is not testable, and it is true that in the presence of unobserv-able characteristics such as cultural norms, two households identical with respect of allthe variables except the treatment assignment may be different in a way unobserved by astatistician. Therefore, we try to minimize the bias in matching households by includingas many variables as possible, including endogenous variables, and by including quadraticterms and interaction terms of those variables.

Table 5 presents the results of estimating a probit model for having a child between1993 and 1997 with a set of variables satisfying the balancing property. Among the de-mographic variables, the age of previous children is an important determinant in additionto the number of previous children to explain a further childbearing. The households thathave more children under five face a higher likelihood of a new birth as the householdhead ages. When there are more children aged 10 to 14, a household is less likely toexperience a birth and the tendency is mitigated as the number of children of 10 to 14increases. When more generations live together, a household has higher chance of havinga birth.

The household head’s characteristics turn out to be an important predictor for newchildbearing. As a household head becomes older, a household is more likely to havea newly born child, and this tendency becomes stronger as the head gets older. Muslimhouseholds are associated with a higher chance of having a birth, whereas those headswho worked in the previous year are associated with a lower chance. The educationallevel of household head is negatively correlated with childbearing only when the head isfemale.



Table 5: Probit estimation of having a new child between 1993 and 1997

Explanatory variables Coefficients Std. ErrorDependent variable: Index for having a new-born child between 1993 and 1997Demographic Urban -0.035 (0.051)characteristics No. of children of ages 0-4 -0.084 (0.134)

No. of children of ages 0-4 sq. -0.013 (0.036)No. of children of ages 0-4 * Head’s age 0.005 * (0.002)No. of children of ages 5-9 0.012 (0.139)No. of children of ages 5-9 sq. 0.017 (0.037)No. of children of ages 5-9 * Head’s age -0.001 (0.003)No. of children of ages 10-14 -0.521 ** (0.125)No. of children of ages 10-14 sq. 0.101 ** (0.027)No. of children of ages 10-14 * Head’s age 0.004 (0.002)No. of adults -0.069 (0.047)No. of generations 0.173 ** (0.051)

Household Head’s age -0.052 ** (0.011)head’s Head’s age sq./100 0.029 ** (0.010)characteristics Head being Muslim 0.309 ** (0.069)

Head worked last year -0.273 ** (0.087)Head worked last year * Head’s female 0.167 (0.148)Head’s years of schooling -0.011 (0.009)Head’s years of schooling * Head’s female -0.053 ** (0.018)Head’s female -0.060 (0.142)

Educational No. of adult men with more than compulsory education -0.013 (0.063)attainment No. of adult women with more than compulsory education 0.139 ** (0.048)

No. of adult men with only compulsory education -0.001 (0.074)No. of adult women with only compulsory education 0.079 (0.048)No. of adult men with less than compulsory education -0.025 (0.081)

Working status No. of adult men who worked last year 0.196 ** (0.055)No. of adult women who worked last year -0.002 (0.035)No. of male children who worked last year -0.110 (0.148)No. of female children who worked last year -0.158 (0.161)Any member worked as a farmer last year 0.069 (0.050)Constant 0.621 * (0.275)

No. of observations 4,694Log Likelihood -2,406.96

Notes: 1) Standard errors are in parentheses.2) * significant at 5%; ** significant at 1%.3) The unit of observation is a household.4) All the explanatory variables denote the values in 1993.5) 179 households with more than one newborn child were dropped.



Among a set of variables representing educational attainment of household membersonly the number of women with more than primary education exhibits a significant as-sociation with having a child. The fact that higher female education is correlated withchildbearing suggests that female education reflects higher income in the context of In-donesia.

Regarding the working status of household members, the only significant variable isthe number of men working, which is positively associated with a new birth. This againsuggests an income effect where higher labor income leads to more children.12

Table 6 presents the effect of a newly born child on expenditure in a household usingthe matching method described in Section 4 for the whole sample. The outcome is thedifference in the household real monthly expenditure per equivalence scale.13 A childis defined as a household member of age below fifteen. When the expenditure per per-son is used as a measure of individual household income, as in the first column, a newlyborn child has a significant negative impact on expenditure per person. Using the nearestneighbor matching method, the households with a new born child between the two wavesexperienced a decrease in consumption per person by 7,600 Rupiah, or 20 percent of con-sumption per person in 1997. This implies a substantial impact of fertility on householdeconomy. Other matching methods produce similar results in terms of magnitude.

When a child is counted as half of an adult, as in the second column, the ATT issignificant and negative, and the magnitude is in the range of 4,200 to 4,800 Rupiah. Thisis around 40% less than the per-capita reduction in expenditure. Although this result isexpected, the level of change is surprising. The cost of a child is calculated to be 52% ofthat of an adult based on the comparison of the baseline household and the household oftwo adults and one child in Table 1. Therefore, the assumption that the weight of a child’sconsumption is one half is not totally unrealistic. The point is that the effect of a newborn child on the individual household consumption can decrease by 40% when a childconsumes a half of what an adult consumes.

12 Strictly speaking there is no theoretical justification for having the number of adults working rather than theproportion of them as an explanatory variable. In principle the number and the proportion convey the sameinformation. However, we are careful in introducing nonlinear variables in the specification because thosevariables seem to be hard to interpret for the marginal effects. Moreover, the specifications with nonlinearvariables tend to make the balancing property unsatisfied. Therefore, we stick to the linear specification.

13 If a birth took place near the time of the survey in 1997, then the consumption expenditure might be measuredbefore the birth (thereby measuring the outcome before the treatment). To address the issue, we conducted thefollowing analysis dropping from the treatment group the households whose newly born child is measured aszero years of age in 1997. The qualitative result remains the same as presented in the Table 6.



Table 6: Effect of fertility on household expenditureper equivalence scale I

Equivalence scale Cons. per Case A Case B Case Cperson

α 1.00 0.50 1.00 0.50θ 1.00 1.00 0.50 0.50

Matching method n(T) n(C)Nearest neighbor 1,135 853 −76.169∗∗ −50.339∗∗ −20.347 13.320

(18.318) (20.700) (39.841) (36.119)Radius 1,135 3,546 −76.025∗∗ −42.159∗∗ −7.361 34.697

(9.710) (12.568) (19.286) (26.586)Kernel 1,135 3,546 −80.243∗∗ −52.637∗∗ −17.531 18.202

(9.888) (13.543) (20.718) (20.501)Stratification 1,135 3,546 −82.023∗∗ −56.011∗∗ −20.560 13.557

(10.585) (13.384) (20.946) (24.550)

Notes: 1) The dependent variable is the difference in real expenditure per equivalent scale between 1993 and1997 (100 Rupiah).2) Standard errors (S.E.) are computed using bootstrap.3) 179 households with more than one newborn child were dropped.4) Treatment is to have a newly born child between 1993 and 1997.5) n(T) and n(C) denote the size of treatment group and controlled group, respectively.6) * significant at 5%; ** significant at 1%.

Next, the role of economy of scale is examined. The third column in Table 6 dealswith a case where the economy of scale parameter is one half with a child consuming asmuch as an adult. The ATT is still negative but not significant. The magnitude of the ATTis in the range of 1,000 to 1,300 Rupiah, which is only around 16% of the reduction inper-capita expenditure. That is, the effect of new childbearing on individual householdconsumption almost disappears when the parameter of economy of scale is one half. Un-der this assumption, the expenditure of a household with three adults increases by 32%compared to a household of two adults at the same level of utility.14 The estimated equiv-alence scale in Table 1 suggests that the expenditure increases by 16% from a householdof two adults to that of three adults. Therefore, the size of one half as a parameter ofeconomy of scale is reasonable.

14 The calculation is based on the comparison of√

2 and√

3.



When both the relative weight of a child and the economy of scale parameter are set tobe one half as in column (4) of Table 6, the ATT becomes positive but insignificant. Theresult is well expected from columns (2) and (3), but it is still surprising that the individualhousehold consumption can increase even after a new birth under a different equivalencescale. When the significance of the estimates in Table 6 is considered, the exercise at leastsuggests that the negative effect of a new birth on individual household consumption candisappear for some range of realistic values of the equivalence scale. 15 16

Next we estimate the effect of a child on household consumption applying the es-timates of the equivalence scale obtained in Table 2. The results are presented in Ta-ble 7. The first column is the case where expenditure per person is used as a measure ofhousehold consumption. The effects of a newly born child on consumption per estimatedequivalence scale in the other columns are all negative although the estimates in column(3) and (4) are not precisely estimated. The magnitudes of the ATT are in the range from20 percent to 65 percent of that estimated when the per-capita expenditure is used as ameasure of consumption. Since the estimates of both equivalence scale parameters in Ta-ble 2 are between 0.5 and 1, the ATT in column (4) of Table 6 serves as an upper boundof the effect of a birth on individual household consumption. On the other hand, whenit is used as a measure of individual consumption, the expenditure per person is likely toexaggerate the effect of a new born child on individual household consumption at least by35%.

15 Given the huge geographical diversity in Indonesia, one might expect a systematic difference between ruraland urban samples due to the differences in relative prices or in the child care cost. However, it is found thatthe patterns of estimates with different equivalence scales in rural and urban samples are similar to that from thetotal sample.

16 One observation is notable. The average treatment effect estimated using nearest neighbor matching methodtends to be greater than the other estimates in both sub-samples. Further, the estimates seem to be ranked in theorder of nearest neighbor matching, radius matching, kernel matching, and stratification matching. However, toour knowledge, no studies are able to verify that there is a systematic difference in the magnitudes of the averagetreatment effect from different matching methods.



Table 7: Effect of fertility on household expenditureper equivalence scale II

Equivalence scale (ES) Cons. per Implied Implied Impliedperson ES 1 ES 2 ES 3

α 1.00 0.65 0.76 0.92θ 1.00 0.90 0.71 0.56

Matching method n(T) n(C)Nearest neighbor 1,135 853 −76.169∗∗ −54.287∗∗ −42.877 −27.683

(18.318) (20.914) (25.913) (32.502)Radius 1,135 3,546 −76.025∗∗ −47.248∗∗ −33.178∗ −15.625

(9.710) (11.270) (16.115) (19.355)Kernel 1,135 3,546 −80.243∗∗ −56.250∗∗ −42.872∗∗ −25.685

(9.888) (14.446) (15.829) (19.121)Stratification 1,135 3,546 −82.023∗∗ −59.196∗∗ −45.903∗∗ −28.721

(10.585) (14.164) (15.759) (20.025)

Notes: 1) The dependent variable is the difference in real expenditure per equivalent scale between 1993 and1997 (100 Rupiah).2) Standard errors (S.E.) are computed using bootstrap.3) 179 households with more than one newborn child were dropped.4) Treatment is to have a newly born child between 1993 and 1997.5) n(T) and n(C) denote the size of treatment group and controlled group, respectively.6) * significant at 5%; ** significant at 1%.

6. Private expenditure on child and intra-household bargaining

The finding of the previous section has implications on the relationship between fertilityand the expenditure on private-public goods. Our results indicate that fertility affectshousehold consumption differently depending on the assumption about the economy ofscale and the relative weight of a child consumption to that of an adult (see Table 6).These findings in turn, imply that the characteristics of household consumption is likelyto affect fertility.17 In particular, the composition of the expenditure on private and publicgoods may reflect the price of having a child. For example, the share of private goodssuch as children’s education in household expenditure can be interpreted as indicating a

17 We owe this section to two anonymous referees, who suggested adding the discussion on fertility, the con-sumption of public-private goods and the intra-household bargaining.



household’s preference for quality of a child. The larger share of educational expenditurea household has, the less likely it is to give birth. We test this hypothesis using a probitmodel as used in Table 5. As a measure of the expenditure on private goods, three vari-ables are used: share of educational expenditure, educational expenditure per person andeducational expenditure per child.18

The results are displayed in Table 8. The basic specification in column (1) showsthe result similar to that of Table 5. Most variables representing household characteristicsand provinces are abbreviated in order to keep the focus on educational expenditure. Totalhousehold expenditure per capita does not exhibit any significant impact on fertility in thesample of all the households, whereas a positive income effect is found for the group ofhouseholds with any child in column (4).

Table 8: The effect of educational expenditure on fertility (Probit Model)

Dependent variable: Index for (1) (2) (3) (4)having a newly-born child All All All HH withbetween 1993 and 1997 childrenShare of educational expenditure - −0.864∗∗ - -

(0.199)log Educational expenditure - - −0.084∗∗ -per person (0.015)log Educational expenditure - - - −0.072∗∗per child (0.015)log Total household expenditure −0.003 0.012 0.059 0.083∗per person (0.035) (0.035) (0.036) (0.042)

No. of observations 4, 694 4, 694 4, 694 3, 480Log Likelihood −2, 421.43 −2, 411.55 −2, 405.13 −1, 880.73

Notes: 1) Standard errors are in parentheses.2) * significant at 5%; ** significant at 1%.3) The unit of observation is a household.4) All the explanatory variables denote the values in 1993.5) 179 households with more than one newborn child were dropped.6) Household characteristics and province dummies are included in all specifications.

18 The educational expenditure as a part of the information on consumption expenditure is measured at thehousehold level, and it is not separated for adults’ education and children’s. Therefore, both results with theeducational expenditure scaled in terms of per capita and per child are presented below.



Column (2) of Table 8 indicates that a household spending more on education is lesslikely to give birth as expected. The implied marginal effect suggests that one percentchange in educational share of expenditure is associated with a quarter percent change inthe probability of giving birth between 1993 and 1997. The absolute level of educationalexpenditure per person also has a negative effect on the likelihood of having a new childas in column (3). Using the educational expenditure per child as a measure of the educa-tional expenditure generates the similar result, where only households with any child areconsidered (column (4)). The changes in probability of giving birth associated with onepercent change in educational expenditure per person and per child are 0.25 percent and0.23 percent, respectively. Therefore, it may be inferred that those households with moreexpenditure on private goods like education are less likely to give birth.

In examining the relationship between fertility and household consumption so far, wehave assumed that a household makes a decision the same way as an individual does.If a household behaves collectively through a bargaining process among its members,the specification for equivalence scale needs to be modified accordingly as suggested byBrowning et al. (2006). Therefore, it would be informative to investigate whether theframework of a unitary household is supported by the same data.

In understanding the relationship between household preference and fertility, the de-mand for private goods may be determined by household members with different prefer-ence and bargaining power (Browning and Chiappori 1998). Specifically, the expenditureon children’s education may be determined by the husband and wife’s collective decision.It is a common conjecture that mothers tend to care more about children’s education, orthat mothers at least have different preference from fathers. Therefore, we focus on in-vestigating the differences in preferences of husband and wife with regard to children’seducation in the context of Indonesia. Using the first wave of the IFLS, the effect of thebargaining power on the educational share of expenditure is examined.

One measure of bargaining power suggested in the literature is nonlabor income(Lundberg et al. 1997; Duflo 2003; Park 2007). In the IFLS, the nonlabor income isconstructed by combining income from pension, scholarship, insurance claim, lottery orgift from family, friends and charity. One limitation is that the percentage of the house-holds with positive nonlabor income is relatively small (17% for husband’s nonlabor in-come, 22% for wife’s at household level). Another measure indicating bargaining poweris premarital assets (Quisumbing and Maluccio 2003; Park 2007). The IFLS has the infor-mation on premarital assets, but there is a substantial attrition (48%) due to the availabilityof premarital assets and the relevant price indices.

In addition, we consider the share of the current asset as a measure of bargainingpower assuming that the individual shares of current assets are exogenously given. Threekinds of assets are used: farm business, non-farm business and house. Because therecorded values of assets are not reliable, the relative ownership of an asset is taken as



a measure of bargaining power. That is, the sum of shares of ownership between husbandand wife is set to be one. A husband’s share of one asset, for example, farm businessshould be interpreted to be valid among households with farm business since it is, in fact,the interaction term between farm business owner dummy and the husband’s share.

As discussed in the literature, we test for unitary versus collective household decisionprocess by investigating whether the individual resources of husband and wife have thesame effect on educational expenditure. Also we examine whether husband’s share ofassets have any impact on the expenditure on education.19

Regarding the dependent variable, there are two sources for educational expenditurein the IFLS. The first is the information on educational expenditure as a part of the house-hold consumption expenditure. The educational expenditure for a household includes theexpenditure on adults’ education as well as children’s education. The second is the ed-ucational expenditure for a particular child. The IFLS has a special section on childrenunder age 14, which was conducted for two randomly chosen children per household.The educational expenditure on a child was collected only for those attending school. Weuse both pieces of information. First, the household educational expenditure per personis used as a measure of educational expenditure with the assumption that the expenditureon adults’ education is negligible. Second, the results using the expenditure on a schoolchild as a measure of educational expenditure in a sample of children are presented.

The results using the data at the household level are presented in Table 9. According tocolumn (1), households with non-farm business tend to spend less on education, whereashouseholds with farm business or house do not exhibit any difference. Husband’s shareof three assets does not have any impact on educational spending. Households with largertotal household expenditure per person tend to spend more on education. One percentincrease in per capita household expenditure is associated with 2.5 percent increase inshare of the expenditure on education. This magnitude is roughly consistent in column(2) and (3). In column (2) the husband’s and wife’s premarital assets have an impact oneducational spending with different direction, but the result does not reject the hypothesisthat they have the same impacts at the conventional level of significance. Husband’sand wife’s nonlabor income does not show any effects significantly different from eachother as in column (3). When the log of educational expenditure per person is used as adependent variable in column (4) to (6), the results remain qualitatively the same.

19 Under the null hypothesis that the resources of a husband and a wife have the same effects, the husband’sshare should have no effect on educational expenditure.



Table 9: Household bargaining and educational spending I

(1) (2) (3) (4) (5) (6)Dependent variable Share of educational log Educational expenditure

expenditure per personFarm business owner 0.003 - - 0.040 - -

(0.004) (0.052)Farm business −0.006 - - −0.080 - -husband’s share (0.009) (0.111)Non-farm business −0.009∗ - - −0.037 - -owner (0.004) (0.046)Non-farm business 0.003 - - 0.022 - -husband’s share (0.008) (0.097)House owner −0.002 - - 0.019 - -

(0.005) (0.064)House husband’s 0.009 - - 0.052 - -share (0.009) (0.107)log Husband’s - −0.001 - - −0.002 -premarital asset(α1) (0.001) (0.009)log Wife’s - 0.002 - - 0.011 -premarital asset(α2) (0.001) (0.012)log Husband’s - - −0.0004 - - −0.003nonlabor income(α3) (0.0010) (0.013)log Wife’s nonlabor - - −0.0001 - - 0.020income(α4) (0.0009) (0.012)log Total household 0.025∗∗ 0.018∗∗ 0.023∗∗ 0.806∗∗ 0.800∗∗ 0.802∗∗expenditure per person (0.004) (0.005) (0.004) (0.039) (0.051) (0.039)No. of observations 4,595 2,412 4,466 4,595 2,412 4,466R2 0.19 0.21 0.19 0.44 0.47 0.44p-value for:

H0 : α1 = α2 - 0.08 - - 0.38 -H0 : α3 = α4 - - 0.80 - - 0.21

Notes: 1) White heteroscedasticity-corrected standard errors are in parentheses.2) * significant at 5%; ** significant at 1%.3) The unit of observation is a household.4) All the explanatory variables denote the values in 1993.5) 179 households with more than one newborn child were dropped.6) Household characteristics and province dummies are included in all specifications.



The results from the child level data are shown in Table 10. In column (1), householdswith non-farm business tend to spend more on children’s education in contrast to column(1) in Table 9. The difference is likely to be due to the selection of samples. Householdswith other assets do not spend on education any differently. Husband’s share of the threeassets do not have any significant effects on children’s educational spending. In column(2) the premarital assets of a husband and a wife do not have any significant impact oneducation nor are their effects different from each other. Nonlabor income of a husbandand a wife in column (3) is associated with a larger spending on education, but the resultdoes not reject the hypothesis that their effects are of the same magnitude.

One limitation of the analysis is that the size of samples vary over the specificationsdue to the availability of data. However, data at both household and child level produceresults in favor of a unitary household model. The results are consistent with Park (2007),who did not find a clear evidence that a parental household bargaining plays an impor-tant role in the decision of expenditure on children’s education using the third wave ofIFLS.20 Therefore, the framework of a unitary household seems to be valid at least inunderstanding the relationship between fertility and household consumption.

Table 10: Household bargaining and educational spending II

(1) (2) (3)Dependent variable:log Educational expenditure for a childFarm business owner −0.087 - -

(0.051)Farm business husband’s share 0.097 - -

(0.106)Non-farm business owner 0.125∗∗ - -

(0.041)Non-farm business husband’s share 0.043 - -

(0.087)House owner −0.073 - -

(0.061)House husband’s share −0.166 - -

(0.099)

20 Park (2007) also found a consistent evidence in favor of a collective household model examining the children’snutritional status, and suggested that the decision process may differ depending on the type of decision.



Table 10: (Continued)

(1) (2) (3)Dependent variable: log Educational expenditurefor a childlog Husband’s premarital assets (α1) - 0.010 -

(0.008)log Wife’s premarital assets (α2) - −0.008 -

(0.009)log Husband’s nonlabor income (α3) - - 0.019

(0.010)log Wife’s nonlabor income (α4) - - 0.015

(0.010)log Total household expenditure 0.376∗∗ 0.351∗∗ 0.395∗∗per person (0.036) (0.046) (0.037)Female 0.021 0.032 0.026

(0.034) (0.042) (0.034)Age 0.113∗∗ 0.122∗∗ 0.109∗∗

(0.009) (0.011) (0.010)

No. of observations 2,665 1,716 2,612R2 0.34 0.35 0.34p-value for:

H0 : α1 = α2 - 0.14 -H0 : α3 = α4 - - 0.81

Notes: 1) White heteroscedasticity-corrected standard errors are in parentheses.2) * significant at 5%; ** significant at 1%.3) The correlation between children from the same household is allowed when the standard errors arecalculated.4) The unit of observation is a child.5) All the explanatory variables denote the values in 1993.6) 179 households with more than one newborn child were dropped.7) Household characteristics and province dummies are included in all specifications.



7. Conclusion

This paper describes the causal effect of fertility on household consumption in Indonesiaat the end of a fertility transition in a unitary household framework. Using the nonpara-metric propensity score matching method, it is found that a newly-born child leads to areduction of consumption by 20 percent within four years, when consumption expendi-ture per person is used as a measure of household consumption. This effect, however, isextremely sensitive to the choice of equivalence scale. Using the estimates of equivalencescale based on the assumption that food share inversely indicates the level of householdwelfare, it is shown that the effect of a child on household consumption is still negativebut the magnitude is in the range from 20 to 65 percent of that obtained when per-capitaconsumption is used as a measure of household consumption. Hence, the results suggestthat the analysis based on the conventional measure of poverty is likely to exaggerate theeffect of fertility on poverty at least because the equivalence scale is not properly takeninto account (e.g. Mason and Lee 2004).

The prevalence of poverty measure based on consumption per person is mainly dueto its practical convenience and the strong assumptions required for alternative measures.Our analysis presents a reasonable and practical way of estimating the equivalence scaleand applying it to the question on the relationship between fertility and poverty. Theincreasing quality of survey data in recent years enables one to incorporate the equivalencescale in the poverty analysis in demographic research at least within a certain country.

The sensitivity of the effect of fertility on household consumption with respect to theequivalence scale generates two theoretical implications. First, it illustrates the impor-tance of separating public from private consumption. The expenditure on private goodssuch as children’s education reflects the price of having a child. Second, our analysis pro-vides an empirical bridge between unitary and collective models of the household. Theresults suggest that the expenditure share of private goods can be used to test for a collec-tive decision making process in a household. Although the data are limited, our findingis that the demand for children’s education does not seem to be affected by the parentalbargaining process.

There are a couple of directions suggested for future research. First, since the choiceof welfare indicator may make a considerable difference in the estimation as Anand andHarris (1994) suggested, it is worthwhile to extend the analysis in the paper by using othermeasures of household welfare including income or body mass index. Secondly, becausepoverty is measured as an aggregate variable in a household in this study, it is not clearhow fertility is going to affect total household income. Therefore, estimating the effect offertility on the elements of household income, such as the working hours or earnings perhour of household members in a structural model, will be a fruitful way of investigatingthe causal relationship between fertility and household welfare.



8. Acknowledgements

The authors would like to thank Fabrizia Mealli, Letizia Mencarini and Steve Pudney fortheir helpful comments. They also thank two anonymous referees who essentially helpedto improve the exposition of the paper. The project was funded under the frameworkof the European Science Foundation (ESF) - European Research Collaborative Program(ERCPS) in the Social Science, by Economic and Social Research Council (award no.RES-000-23-0462), the Italian National Research Council (Posiz. 117.14), and the Aus-trian Science Foundation (FWF, contract no. P16903-G05).



References

Anand, S. and Harris, C. J. (1994). Choosing a welfare indicator. The American EconomicReview 84(2): 226–231.

Angrist, J. and Evans, W. (1998). Children and their parents’ labor supply: Evidence fromexogenous variation in family size. The American Economic Review 88(3): 450–477.

Banks, J. and Johnson, P. (1994). Equivalence scale relativities revisited. EconomicJournal 104(425): 883–890. doi: 10.2307/2234982.

Becker, S. O. and Ichino, A. (2002). Estimation of average treatment effects based onpropensity scores. The Stata Journal 2: 358–377.

Birdsall, N., Kelley, A. C., and Sinding, S. W. (2001). Population matters: Demographicchange, economic growth, and poverty in the developing world. Oxford UniversityPress.

Blundell, R. and Lewbel, A. (1991). The information content of equivalence scales.Journal of Econometrics 50: 49–68. doi: 10.1016/0304-4076(91)90089-V.

Browning, M. and Chiappori, P. A. (1998). Efficient intra-household allocations: Ageneral characterization and empirical tests. Econometrica 66(6): 1241–1278.

Browning, M., Chiappori, P. A., and Lewbel, A. (2006). Estimating consump-tion economies of scale, adult equivalence scales, and household bargaining power.Department of economics discussion paper series 289. Oxford: University of Oxford.doi: 10.2307/2999616.

Dawid, A. P. (1979). Conditional independence in statistical theory. Journal of the RoyalStatistical Society Series B 41: 1–31.

De Santis, G. and Maltagliani, M. (2003). Equivalence scales: A fresh look at anold problem. theory and empirical evidence. In: Camilo, D. and Guido, F. (eds.).Household behaviour, equivalence scales, welfare and poverty, pp. 29–54. New York:Physica-Verlag Heidelberg.

Deaton, A. and Paxson, C. (1998). Economies of scale, household size, and the demandfor food. Journal of Political Economy 106(5): 897–930. doi: 10.1086/250035.

Deaton, A. S. and Muellbauer, J. (1986). On measuring child costs: With applications topoor countries. Journal of Political Economy 94(4): 720–744. doi: 10.1086/261405.

Duflo, E. (2003). Grandmothers and granddaughters: Old-age pensions and intra-household allocation in South Africa. World Bank Economic Review 17(1): 1–25.doi: 10.1093/wber/lhg013.


http://dx.doi.org/10.2307/2234982

http://dx.doi.org/10.1016/0304-4076(91)90089-V

http://dx.doi.org/10.2307/2999616

http://dx.doi.org/10.1086/250035

http://dx.doi.org/10.1086/261405

http://dx.doi.org/10.1093/wber/lhg013


Frankenberg, E. and Karoly, L. (1995). The 1993 Indonesian family life survey: Overviewand field report. Field report. Santa Monica: RAND.

Frankenberg, E., Smith, J. P., and Thomas, D. (2003). Economic shocks, wealth andwelfare. Journal of Human Resources 38(2): 280–321.

Frankenberg, E. and Thomas, D. (2000). The Indonesia family life survey (IFLS): Studydesign and results from waves 1 and 2. (dru-2238/1-nia/nichd). Santa Monica: RAND.

Heckman, J., Ichimura, H., and Todd, P. E. (1998). Matching as an econometric evaluationestimator. Review of Economic Studies 65: 261–294. doi: 10.1111/1467-937X.00044.

Heckman, J. J. and Robb, R. (1985). Alternative methods for evaluating theimpact of interventions: An overview. Journal of Econometrics 30(1): 239–267.doi: 10.1016/0304-4076(85)90139-3.

Koulovatianos, C., Schröder, C., and Schmidt, U. (2005). On the incomedependence of equivalence scales. Journal of Public Economics 89: 967–996.doi: 10.1016/j.jpubeco.2004.09.005.

Lanjouw, P. and Ravallion, M. (1995). Poverty and household size. Economic Journal105(433): 1415–1434. doi: 10.2307/2235108.

Lipton, M. (1998). Successes in anti poverty. Geneva: International Labor Organization.

Livi-Bacci, M. and De Santis, G. (1998). Population and poverty in developing world.Oxford: Clarendon Press.

Lundberg, S. J., Pollak, R. A., and Wales, T. J. (1997). Do husbands and wives pool theirresources? evidence from the United Kingdom child benefit. The Journal of HumanResources 32(3): 463–480. doi: 10.2307/146179.

Mason, A. and Lee, S. H. (2004). The demographic dividend and poverty reduction. Paperpresented in seminar on the relevance of population aspects for the achievement of themillennium development goals, November 17-19 2004. New York, U.S.A.

Mattei, A. (2004). Estimating causal effects in experimental and observational studiessuffering from missing data. Ph.d thesis, Firenze: Applied Statistics, Dipartimento diStatistica "G. Parenti", Universita degli Studidi Firenze.

McNicoll, G. (1997). Population and poverty: A review and restatement. Policy researchdivision working papers 105. New York: Population Council.

Merrick, T. (2001). Population and poverty in households: A review of reviews. In:Birdsall, N., Kelley, A. C., and Sinding, S. W. (eds.). Population matters: Demographicchange, economic growth, and poverty in the developing world, pp. 201–212. Oxford:


http://dx.doi.org/10.1111/1467-937X.00044

http://dx.doi.org/10.1016/0304-4076(85)90139-3

http://dx.doi.org/10.1016/j.jpubeco.2004.09.005

http://dx.doi.org/10.2307/2235108

http://dx.doi.org/10.2307/146179


Oxford University Press.

Moav, O. (2005). Cheap children and the persistence of poverty. The Economic Journal115(500): 88–110. doi: 10.1111/j.1468-0297.2004.00961.x.

Park, C. (2007). Marriage market, parents’ bargaining powers, and children’s nutri-tion and education. Oxford Bulletin of Economics and Statistics 69(6): 773–793.doi: 10.1111/j.1468-0084.2007.00479.x.

Quisumbing, A. R. and Maluccio, J. A. (2003). Resources at marriage and intrahouseholdallocation: Evidence from Bangladesh, Ethiopia, Indonesia, and South Africa. OxfordBulletin of Economics and Statistics 65(3): 283–327. doi: 10.1111/1468-0084.t01-1-00052.

Ravallion, M. (1996). Issues in measuring and modelling poverty. The Economic Journal106(438): 1328–1343. doi: 10.2307/2235525.

Rosenbaum, P. T. and Rubin, D. B. (1983). The central role of propensity score in obser-vational studies. Biometrika 70: 41–55. doi: 10.1093/biomet/70.1.41.

Rosenzweig, M. and Schultz, T. P. (1985). The demand for and supply of births: Fertilityand its life cycle consequences. The American Economic Review 75(5): 992–1015.

Rosenzweig, M. and Wolpin, K. (1980). Testing the quantity-quality fertility model:The use of twins as a natural experiment. Econometrica 48(1): 227–240.doi: 10.2307/1912026.

Rosenzweig, M. and Wolpin, K. (2000). Natural “natural experiments” in economics.Journal of Economic Literature 38(4): 827–874.

Schoumaker, B. and Tabutin, D. (1999). Relationship between poverty and fertilityin Southern countries: Knowledge, methodology and cases. Department of scienceof population and development working paper 2. Louvain-la-Neuve: UniversitaCatholique de Louvain.


http://dx.doi.org/10.1111/j.1468-0297.2004.00961.x

http://dx.doi.org/10.1111/j.1468-0084.2007.00479.x

http://dx.doi.org/10.1111/1468-0084.t01-1-00052

http://dx.doi.org/10.1111/1468-0084.t01-1-00052

http://dx.doi.org/10.2307/2235525

http://dx.doi.org/10.1093/biomet/70.1.41

http://dx.doi.org/10.2307/1912026


Appendix: Estimating Equivalence Scale using Engel method

Calculating the equivalence scale requires an assumption on the measure of welfare. Twomethods of constructing a proxy of welfare are widely used in the literature. The En-gel method assumes that the food share of expenditure inversely indicates the level ofhousehold welfare. The Rothbarth method assumes that the consumption of adult goodscorrectly indicates the level of adults’ welfare in a household. With the limitation onthe detail of consumption measure in the IFLS, we take the Engel method in estimatingequivalence scale.

The food share of expenditure is considered as a simple linear function of the house-hold expenditure per person and the demographic characteristics as in Deaton and Muell-bauer (1986).

wf = α + β ln(x

n) +

J∑

j=1

γjnj + ε (11)

where wf is the food share of household, x is the total household expenditure, n denotesthe number of household members, nj is a set of the variables indicating the demographicstructure of a household, and ε represents a residual. The results are presented in Table11. The ordinary least squared (OLS) estimation suggests that ten percent increase in ex-penditure per person is associated with a decrease of 0.013 in the food share. Controllingfor the expenditure per person, an additional adult or child are associated with a lowerfood share, and the magnitude of the change is larger for an adult. Since there exists ahuge geographic diversity in Indonesia, we run the community fixed-effects estimation inorder to control for the variation in prices or infrastructure at the community level. Theresults, shown in column (2), are similar to those from the OLS estimation.

With the estimates of the Engel curve, we calculate the equivalence scale by com-paring two households with different demographic structures (i.e. household composi-tion). We denote the total expenditure and the demographic characteristics of the baselinehousehold of two adults by x0 and n0, respectively. We are interested in the amountof expenditure, xc, that would give the same level of food share to a household with ademographic structure, nc, as that of the baseline household.

α + β ln(xc

n) +

J∑

j=1

γjncj = α + β ln(

x0

n) +

J∑

j=1

γjn0j (12)



The ratio of xc to x0, or the equivalence scale, E(c, 0), is obtained by solving equa-tion (12) for xc

x0.

E(c, 0) =xc

x0=

nc

n0/exp

(J∑

j=1

(γj

β)(nc

j − n0j )

)(13)

Table 11: Engel Curve Estimation

(1) (2)OLS Comm FE

Dependent variable: Food share of expenditurelog x

n −0.127∗∗ −0.119∗∗(0.003) (0.004)

No. of adults −0.032∗∗ −0.027∗∗(0.002) (0.002)

No. of children −0.022∗∗ −0.026∗∗(0.002) (0.002)

Constant 1.462∗∗ 1.407∗∗(0.021) (0.024)

No. of observations 4,852 4,852No. of communities 311R2 0.27 0.21

Notes: 1) Standard errors in parentheses.2) * significant at 5%; ** significant at 1%.


does fertility decrease household consumption? an analysis ... · in addition to joint causation,...

Documents